Published OnlineFirst August 13, 2013; DOI: 10.1158/1078-0432.CCR-12-3863
Clinical Cancer Human Cancer Biology Research
Novel Clinically Relevant Genes in Gastrointestinal Stromal Tumors Identified by Exome Sequencing
Sebastian F. Schoppmann1, Ursula Vinatzer1, Niko Popitsch5, Martina Mittlbock€ 2, Sandra Liebmann-Reindl1, Gerd Jomrich1, Berthold Streubel3, and Peter Birner4
Abstract Purpose: Chromosomal gains and losses resulting in altered gene dosage are known to be recurrent in gastrointestinal stromal tumors (GIST). The aim of our study was the identification of clinical relevant genes in these candidate regions. Material and Methods: A cohort of 174 GIST was investigated using DNA array (n ¼ 29), FISH (n ¼ 125), exome sequencing (n ¼ 13), and immunohistochemistry (n ¼ 145). Results: Array analysis revealed recurrent copy number variations (CNVs) of chromosomal arms 1p, 1q, 3p, 4q, 5q, 7p, 11q, 12p, 13q, 14q, 15q, and 22q. FISH studies of these CNVs showed that relative loss of 1p was associated with shorter disease-free survival (DFS). Analysis of exome sequencing concentrating on target regions showing recurrent CNVs revealed a median number of 3,404 (range 1,641–13,602) variants (SNPs, insertions, deletions) in each tumor minus paired blood sample; variants in at least three samples were observed in 37 genes. After further analysis, target genes were reduced to 10 in addition to KIT and PDGFRA. Immunohistochemical investigation showed that expression of SYNE2 and DIAPH1 was associated with shorter DFS, expression of RAD54L2 with shorter and expression of KIT with longer overall survival. Conclusion: Using a novel approach combining DNA arrays, exome sequencing, and immunohis- tochemistry, we were able to identify 10 target genes in GIST, of which three showed hithero unknown clinical relevance. Because the identified target genes SYNE2, MAPK8IP2, and DIAPH1 have been shown to be involved in MAP kinase signaling, our data further indicate the important role of this pathway in GIST. Clin Cancer Res; 19(19); 5329–39. 2013 AACR.
Introduction known to inhibit both KIT and PDGFRA receptors and Gastrointestinal stromal tumors (GIST) are the most used in the treatment of recurrent and metastatic GIST or common mesenchymal tumors of the gastrointestinal tract GIST with high risk of progression (5). Effectiveness of (1). They are thought to arise from Cajal cells or their imatinib mesylate depends on the mutational status of KIT PDGFRA precursors and characteristically harbor gain of function and (3). In contrast to metastatic GIST, mutations in KIT leading to constitutive activation of the radical surgery seems to be the best treatment option for KIT receptor (2). localized tumors (6). The recurrence rate after radical An alternative mutation in a related tyrosine kinase, surgery seems to depend mainly on tumor localization, PDGFRA, is found in 35% of KIT mutation negative GIST size, and mitotic activity and ranges between 5% and (3, 4). Imatinib mesylate is a molecularly targeted drug 75% with a poor clinical outcome in relapsed patients (7, 8). Current classifications take into account tumor size, mitotic rate, and tumor location but do not include Authors' Affiliations: 1Department of Surgery; 2Center for Medical Sta- mutational data nor protein expressions of tumor cells 3 tistics, Informatics, and Intelligent Systems; Department of Obstetrics and (9–11). Gynecology and Core Unit Next Generation Sequencing; 4Clinical Institute of Pathology, Medical University of Vienna; and 5Center for Integrative The discovery of KIT and PDGFRA activating mutations in Bioinformatics Vienna (CIBIV), Max F Perutz Laboratories, University of the majority of GIST represents a significant progress in Vienna & Medical University of Vienna, & Faculty of Computer Science, University of Vienna, Vienna, Austria understanding their biological behavior. Although further molecular mechanisms underlying the development and Note: Supplementary data for this article are available at Clinical Cancer Research Online (http://clincancerres.aacrjournals.org/). progression of GIST are not fully understood, they are nec- essary due to the wide range of clinical behavior. Chromo- Corresponding Author: Berthold Streubel, Department of Obstetrics and Gynecology, Medical University of Vienna, Wahringer€ Gurtel€ 18-20, A-1090 somal gains and losses resulting in altered gene dosage are Vienna, Austria. Phone: 43-1404002821; Fax: 43-1404002862; E-mail: known to be recurrent in GIST and believed to have a role [email protected] in the molecular pathogenesis of these tumors (12–19). doi: 10.1158/1078-0432.CCR-12-3863 Nevertheless, the target genes remain to be identified within 2013 American Association for Cancer Research. these regions.
www.aacrjournals.org 5329
Downloaded from clincancerres.aacrjournals.org on October 3, 2021. © 2013 American Association for Cancer Research. Published OnlineFirst August 13, 2013; DOI: 10.1158/1078-0432.CCR-12-3863
Schoppmann et al.
Microarray analysis Translational Relevance High-quality genomic DNA was obtained from 29 fresh- Gastrointestinal stromal tumors (GIST) are the most frozen tumor samples using the DNeasy Blood & Tissue Kit common mesenchymal tumors of the gastrointestinal (Qiagen) and subjected to microarray analysis using the tract, characterized by uncertain clinical behavior. In commercially available Affymetrix Genome-Wide Human addition to KIT, there is a strong need for further SNP Array 6.0 (Affymetrix Inc.) following the protocols therapeutic targets in GIST. Using a technical approach provided by the manufacturer. Data analysis was conducted combining DNA array analysis, next generation exome with Affymetrix Genotyping Console 3.0.1 using the Bird- sequencing, and immunohistochemistry, we were able seed Algorithm and Affymetrix Chromosome Analysis Suite to identify 10 novel, frequently mutated genes in GIST, 1.01 at a resolution of 500 kb. Genome annotations applied of which 3 showed hithero unknown clinical rele- in data analysis referred to the human reference assembly vance. Because a part of the identified target genes has GRCh37/hg19 as provided by the Affymetrix annotation file been shown to be involved in MAP kinase signaling, release na31. The reference model file used for data nor- our data further indicate the important role of this malization with GTC was generated from 39 healthy control pathway in GIST. So the inclusion of MAP kinase individuals. CNVs showing an overlap greater than 80% pathway parameters in future clinical studies might with benign CNVs of the Database of Genomic Variants be of potential benefit for patients as selective inhibi- (http://projects.tcag.ca/variation/) were excluded from our tors are available. analysis. Somatic CNV status of matching tumor-blood/normal tissue samples was used to confirm gains and losses using TaqMan real-time PCR or FISH. The aim of our study was the identification of putative prognostic markers within the regions with recurrent FISH analysis gains or losses. We measured copy number variations Tissue microarrays from 125 GIST paraffin-embedded (CNVs) and defined exact chromosomal breakpoints for samples were used for FISH. Screening for CNVs was con- the most common alterations in GIST. CNVs were screened ducted using commercially available probes for TP73 (chro- in a large single-center cohort of GIST and correlated with mosomal band 1p36), ABL2 (1q25), CCND1 (11q13), clinical outcome. Exome sequencing identified mutated DLEU (13q14), IGH (14q32), SNRPN (15q11), and CLTCL1 candidate genes in the regions of interest, and protein (22q11.2; Abbott and Kreatech). FISH procedures and anal- products of identified target genes were investigated immu- yses were conducted according to standard protocols. nohistochemically in our large single-center cohort of GIST. Immunohistochemical analysis Tissue microarrays containing 145 GIST samples were used for evaluation of protein expression. Supplementary Patients and Methods Table S2 summarizes the antibodies used. Patients Immunohistochemical analysis was conducted using a We studied a total of 174 cases of GIST treated at the Benchmark Ultra Immunostainer (Ventana), except for exp- Medical University of Vienna between August 1992 and ression of RB, where a DAKO autostainer (DAKO) was used. February 2011 in this retrospective observational study. A specimen was considered to be positive if the vast majority Clinical data and follow-up were available for 145 of 174 of cells (>80%) showed distinct staining. The number of cases. All cases were restaged according to UICC TNM cases differed between antibodies due to the use of tissue classification of malignant tumors 7th edition and risk microarrays, and for investigation of FLT4 and AP1B1 the was evaluated according to Fletcher and Miettinen (9– blocks had to be recut, resulting in the loss of several spots. 11). As this cohort of patients has been used in several previous studies, clinical data in correlation with known Exome studies risk factors have been reported previously, as well as the Thirteen of the 29 GIST of the microarray study were results of the sequence analysis about mutations of KIT subjected to high-throughput sequencing. In all these cases, (exons 9, 11, 13, and 17) and PDGFRA (exons 12 and 18; matched control DNA obtained from the peripheral blood refs. 20–22). Tissue microarrays were established for or normal gastric tissue of patients were sequenced in immunohistochemical screening in all 145 cases. FISH parallel. Exome enrichment was conducted using the Tru- was conducted on tissue microarrays of paraffin-embed- Seq sample prep (Illumina). Sequencing was conducted on ded tissue in 125 of 145 cases. Apart from the 145 cases a HiSeq 2000 (Illumina). All datasets were aligned with with clinical follow-up data and paraffin-embedded tis- BWA v 0.6.1-r104 against the human g1k v37 genome using sue available, further 29 GIST cases were identified with standard parameters. We conducted INDEL realignment fresh-frozen tumor material suitable for microarray anal- with GATK v1.4, removed PCR duplicates using Picard ysis and next generation sequencing. Tumor purity was at v1.56 and recalibrated per-base quality values using GATK. least 90% in samples used in all 29 cases. Institutional Variants were then called and filtered using 2 independent review board approval was obtained. pipelines based on GATK’s UnifiedGenotyper (here SNPs
5330 Clin Cancer Res; 19(19) October 1, 2013 Clinical Cancer Research
Downloaded from clincancerres.aacrjournals.org on October 3, 2021. © 2013 American Association for Cancer Research. Published OnlineFirst August 13, 2013; DOI: 10.1158/1078-0432.CCR-12-3863
Novel Clinically Relevant Genes in GIST
and INDELs were filtered separately) and SAMTOOLS, tion. All patients underwent initial surgical treatment; 25 respectively. The variant sets were then merged and we patients received adjuvant imatinib mesylate. Median selected all variants from the matched normal-tumor pairs observation time was 37 3 (SE) months. Although 7 that were either unique in the tumor sample or changed patients showed advanced stage disease already at time of their predicted genotype from being heterozygous in the diagnosis, and no complete surgical removal of the GIST normal sample to being homozygous in the cancer sample. was possible, 24 patients experienced recurrent disease, and We annotated these candidate variants using SnpEff 2.0.4 12 died. RC1 and predicted the possible effect of a variant (e.g., silent/nonsilent/nonsense mutation) using the NCBI Ref- Recurrent chromosomal gains and losses erence Sequence (RefSeq) gene annotations. We further The microarray technology was conducted to define determined whether a variant was already contained in the recurrent chromosomal regions and to determine exact dbSNP TSI (Toscani in Italia) dataset. Candidate variants breakpoints in 29 GIST tumors. Although the microarray were further annotated using polyphen2 and finally man- platform used allows for the detection of CNVs down to 50 ually filtered and inspected using IGV. Genes were com- kb or smaller, we aimed to assess the minimal region of pared with PubMed, Gene database, and the CancerGenes overlap of recurrent genomic regions that are known to be database for gene selection and priorization. Variants in much larger. Therefore, we used a resolution of 500 kb and candidate genes were confirmed by Sanger sequencing in all found 114 CNVs (43 gains and 71 losses) in the 29 tumor tested cases. samples altogether (median number of CNVs/sample). We compared the CNVs between the 29 cases and defined Statistical analysis rearrangements as recurrent when observed in at least 3 Categorical data were described with absolute and rela- cases. Eighty-five of the 114 CNVs (75%) were located tive frequencies. Corresponding 95% confidence intervals within recurrent rearrangements and included losses of n ¼ n ¼ n ¼ (95% CI) were calculated according to the method of chromosomal arms 1p ( 13), 3p ( 4), 13q ( n ¼ n ¼ n ¼ Wilson. Group differences were tested by x2 test or by Fisher 5), 14q ( 17), 15q ( 7), and 22q ( 11) as well as n ¼ n ¼ n ¼ exact test in the case of sparse data. Continuous data are gains of chromosomal arms 1q ( 3), 4q ( 6), 5q ( n ¼ n ¼ n ¼ described with median, minimum, and maximum. Differ- 8), 7p ( 3), 11q ( 4), and 12p ( 3; Table 2). ences between 2 groups for continuous and ordinal data CNVs were confirmed in all samples by FISH and TaqMan were tested by Mann–Whitney U test. assays (1p: Hs05732825_cn; 3p: Hs01499513_cn; 5q: Disease-free survival (DFS) was defined from the day of Hs02944177_cn; 11q: Hs06329804_cn; 13q:Hs07026395_cn; surgery until first evidence of progression of disease if 14q:Hs02834878_cn; 15q: Hs05354966_cn; 22q: complete surgical resection was possible. Patients showing Hs01356996_cn). advanced disease at time of initial diagnosis were excluded A detailed list of CNVs is provided in Table 2. In conclu- from analysis of DFS. Overall survival (OS) was defined as sion, we found that the majority of chromosomal imbal- the time from primary surgery to patients’ death. Survival ances were recurrent, indicating that candidate genes rele- data were graphically presented by Kaplan–Meier graphs vant for GIST pathogenesis are located in these regions. and/or described by survival at predefined time points. Group differences are tested by the log-rank test. Cox Gains and losses in correlation with clinico-pathologic regression models were used to estimate the effect of the data expression of proteins adjusted for patients’ age (<60 vs. In a next step, we investigated if recurrent CNVs are of 60) and risk according to Fletcher’s score. Group differ- prognostic relevance. Seven of the 12 regions (1p, 1q, 11q, ences are quantified with HR and corresponding 95% CIs. 13q, 14q, 15q, and 22q) were investigated on tissue micro- Proportional hazards assumptions were graphically che- arrays containing 125 GIST tumors. About the remaining 5 cked by log-minus-log plots in stratified Cox regression regions, commercially available probes did not provide models for every variable in the Cox model separately. reliable signals on the tissue microarrays. A two-tailed P-value of 0.05 was considered as signif- Results about frequencies of losses and gains are sum- icant. SPSS 20.0 (IBM) was used for all calculations. For marized in Supplementary Table S1 and were comparable exploratory purposes, all P values together were also adjust- to DNA microarrays. CNV status of investigated regions was ed to result in a false discovery rate (FDR) of 0.05 as correlated with clinical risk factor, tumor localization, DFS, KIT PDGFRA described by Benjamini and Hochberg (23). Significant OS, and and mutations status. results after FDR-adjustment are marked by "a". FISH for chromosome 1p was successful in 119 samples. Loss of 1p (Fig. 1) was seen in 11 samples (9.2%; 95% CI, 5.2–15.8%), in 5 additional cases, a chromosomal imbal- Results ance between 1p and 1q was observed (5 1p/1qþ). When Clinical characteristics of patients investigating the 16 cases (13.4%; 95% CI, 8.5–20.7%) with Table 1 summarizes the clinical data of the 145 patients relative imbalance of 1p against all other cases, no associ- with available clinical and follow up data. In brief, 94 GIST ation with clinical risk factor was seen either, but these cases (64.8%) were of gastric localization (64.8%), 92 (63.4%) showed a significantly shorter DFS (5 years DFS rate: 40.4% showed a KIT mutation, and 21 (14.5%) a PDGFRA muta- 17% SE vs. 84.6% 4.4% SE, P ¼ 0.012, log-rank test) but
www.aacrjournals.org Clin Cancer Res; 19(19) October 1, 2013 5331
Downloaded from clincancerres.aacrjournals.org on October 3, 2021. © 2013 American Association for Cancer Research. Published OnlineFirst August 13, 2013; DOI: 10.1158/1078-0432.CCR-12-3863
Schoppmann et al.
Table 1. Clinico-pathologic data of 145 GIST patients included into this study
Number of 5 Years disease-free 5 years overall Factor cases survival rate SE survival rate SE Localization P < 0.001a P > 0.05 Gastric 94 (64.8%) 89.8 4.3% – Duodenum 6 (4.1%) 100% – Small intestine 27 (18.6%) 62.2 10.1% – Rectum 3 (2.1%) 100% – Other 15 (10.3%) 45.1 14.9% – Risk Fletcher P ¼ 0.001a P < 0.001a Very low 16 (11%) 87.5 11.7% 100% Low 50 (34.5%) 93.9 4.2% 95.5 4.4% Intermediate 32 (22.1%) 79 8.6% 96.4 3.5% High 47 (32.4%) 54.8 9.7% 73.8 7.8% Risk Miettinen P < 0.001a P < 0.001a None 18 (12.4%) 90 9.5% 100% Very low 38 (26.2%) 96.3 3.6% 93.8 6.1% Low 28 (19.3%) 77 10.7% 95.8 4.1% Moderate 17 (11.7%) 84.6 10% 100% High 29 (20%) 59.9 12.4% 65.5 9.7% Insufficient data 15 (10.3%) 45.1 14.9% 83.3 15.2% Staging UICC P < 0.00a P > 0.05 pT1 19 (13.1%) 90 9.5% – pT2 65 (44.8%) 84.5 6.2% – pT3 40 (27.6%) 73.6 8.9% – pT4 21 (14.5%) 50.5 12.9% – Mitotic rate UICC P ¼ 0.001a P < 0.001a Low 96 (66.2%) 87.1 4.5% 96.7 2.4% High 49 (33.8%) 58.9 8.9% 75.8 7.1% Synchronous metastases P ¼ 0.003a P < 0.001a No 131 (90.3%) 80.7 4.3% 94.6 2.7% Yes 14 (9.7%) 47.3 19% 51.3 12.5% KIT mutation P > 0.05 P > 0.05 No 53 (36.6%) –– Yes 92 (63.4%) –– PDGFRA mutation P > 0.05 P > 0.05 No 124 (85.5%) –– Yes 21 (14.5%) ––
aSignificant results after FDR adjustment.
not OS (Fig. 2). A significant association of relative loss of 1p of 18 (44.4%: 95% CI, 24.6–66.3%) small intestinal GIST]. with localization of the GIST was found (P ¼ 0.001a, x2 Deletion of 22q correlated with the presence of KIT muta- test). So of 75 gastric GIST, only 4 (5.3%; 95% CI, 5.2– tions. Although in GIST without KIT mutation (n ¼ 28), 33.4%) showed relative loss of 1p, but 9 of 28 (32.1%; 95% only 3 cases (10.7%; 95% CI, 3.7–27.2%) showed a 22q CI, 17.9–50.7%) of small intestinal GIST. deletion, there were 15 of 47 (32%; 95% CI, 20.4–46.2%) FISH for 15q was successful in only 73 samples, and a GIST with KIT mutation (P ¼ 0.038, x2 test). significant association of 15q deletion with tumor locali- zation was seen (P < 0.001a, Fisher exact test), because only Candidate genes within gains and losses 2 of 46 gastric (4.4%; 95% CI, 1.2–14.5%) GIST showed In a next step, we compared the recurrent CNVs that were 15q deletion, but 8 of 17 (47.1%; 95% CI, 26.2–69%) small observed in at least 3 cases (i.e., >10% of total cases) and intestinal GIST. defined minimal regions of overlap shared by all aberrant FISH for 22q was successful in only 75 GIST, and 22q cases. The results are listed in Table 2. The size of the deletion correlated with gastric localization [P ¼ 0.01, x2 minimal regions of interest and consecutively the number test; 7 of 48 (14.6%: 95% CI, 7.3–27.2%) gastric GIST, vs. 8 of genes were too large to identify relevant candidates for the
5332 Clin Cancer Res; 19(19) October 1, 2013 Clinical Cancer Research
Downloaded from clincancerres.aacrjournals.org on October 3, 2021. © 2013 American Association for Cancer Research. Published OnlineFirst August 13, 2013; DOI: 10.1158/1078-0432.CCR-12-3863
Novel Clinically Relevant Genes in GIST
Table 2. Copy number variations assessed by DNA array (n ¼ 29)
Chromosome arm Type of CNV Cases (n/%; 95% CI) ROI Start bp ROI End bp Size (Mb) 1p Loss 13 (45%; 95% CI, 28–63%) 5657217 16851502 11.2 3p Loss 4 (19%; 95% CI, 6–31%) 47478496 50156251 2.7 13q Loss 5 (17%; 95% CI, 8–35%) 45307552 49798504 4.5b 14q Lossa 17 (59%; 95% CI, 41–75%) 26370859 28652135 2.3 41327139 46436189 5.1 54142190 106897379 52.8 15q Loss 7 (24%; 95% CI, 12–42%) 49340635 56142902 6.8 22q Loss 11 (38%; 95% CI, 23–56%) 28772816 51134186 22.4 1q Gain 3 (10%; 95% CI, 5–33%) 118649839 249224376 232.4 4q Gain 6 (21%; 95% CI, 10–38%) 52920476 56180478 3.3c 5q Gain 8 (28%; 95% CI, 15–46%) 120645755 180645101 60.0 7p Gain 3 (10%; 95% CI, 5–33%) 7058423 62706404 55.7 11q Gain 4 (19%; 95% CI, 6–31%) 78846125 134944770 56.1 12p Gain 3 (10%; 95% CI, 5–33%) 479074 15635796 15.2
ROI (region of interest) denotes minimal region of overlap, genome annotations applied in data analysis refer to the human reference assembly GRCh37/hg19. aThe minimal region of overlap on chromosome 14 comprised 3 regions with a total size of 60.2 Mb. bRB1 is located within the minimal region of overlap. cKIT and PDGFRA are located within the minimal region of overlap, respectively. majority. Interestingly, the most frequent gain comprised a 37 genes with the Cancer Genes Database, the Gene Data- small region on chromosome 4 that included KIT and base, and GIST publications. Results are listed in Table 3. PDGFRA. Furthermore, RB1, which was recently reported This strategy reduced the number of candidate genes to 3 or to be a strong predictor of clinical outcome in GIST, was less per region. As expected KIT, PDGFRA, and RB1 were located in the deleted small region of overlap on chromo- among these cases. In addition, we searched for potential some 13. We concluded that our strategy identified success- interactions between the 37 genes and found positive fully the 2 most prominent genes in GIST, that is KIT and matches between DIAPH1 and AP1B as well as MAPK8IP2 PDGFRA, within our regions of interest. Nevertheless, the and SYNE2. candidate remained elusive for the other regions. Protein expression in correlation with risk factors Exome sequencing for the identification of novel Eleven candidate genes were selected for immunohisto- candidate genes chemical evaluation (Table 4 and Supplementary Table Exome sequencing was conducted to identify putative S2). Table 4 shows the number of GIST samples investigat- novel candidate genes within the defined regions of interest ed, the primary staining pattern, and the number of positive as outlined in Table 2. We compared tumors with the cases for each antibody. Figure 1 gives samples of immu- matched constitutional genomes of 13 patients. The medi- nostaining. Gene mutation status has been shown to be an overall number of variants/exome was 50,625 (range linked to protein expression in a variety of human genes, but 38,365–58,258) per exome (Supplementary Table S3), the because protein overexpression might indicate loss of func- median number of variants (SNPs, insertions, deletions) tion (e.g., P53; ref. 24) as well as gain of function (KIT, ALK; present in tumor minus paired blood sample was 3,404 refs. 25, 26), protein expression was scored as positive (range 1,641–13,602; Supplementary Table S4). Intergenic versus negative without any hypothesis about the type of and intronic variants were excluded and annotation by altered gene function. snpEff reduced the average number from 3,404 to 1,764 Expression of RBM5, RB, UBPY/USP8, GTSE1, MAPK8IP2, (range 1,012–7,300). These variants included the switch of APIB1, and FLT4 was not associated with risk according to heterozygous mutations to homozygous mutations. Var- Fletcher, Miettinnen, or UICC staging and grading, DFS or iants were considered only when they were stop mutations, OS or metastases at time of diagnosis (Table 4). RAD54L2 splice site mutations, frame shift mutations, or missense expression was associated with higher risk according to mutations with potential pathogenic at in silico prediction. Fletcher (median high vs. intermediate risk; P ¼ 0.033; In a next step we looked for genes with recurrent variants in Mann–Whitney U test) and higher UICC staging (median at least 3 patients. Recurrent variants were found in 37 genes pT3 vs. pT2; P ¼ 0.032; Mann–Whitney U test). Expression of and listed in Table 3. Sanger sequencing confirmed the RAD45L2 was associated with shorter OS (5 years OS rate: variants in all tested cases. Furthermore, we compared the 80.8% 12.2% SE vs. 93.7% 3.2% SE, P ¼ 0.042, log-rank
www.aacrjournals.org Clin Cancer Res; 19(19) October 1, 2013 5333
Downloaded from clincancerres.aacrjournals.org on October 3, 2021. © 2013 American Association for Cancer Research. Published OnlineFirst August 13, 2013; DOI: 10.1158/1078-0432.CCR-12-3863
Schoppmann et al.